Hello everyone, I have some questions concerning the implementation of pymc3 in statistical forecasting and numerical integration. Here is the problem: I want to construct a hierarchical model with Chinese provinces as different levels, and also with some numerical integration during the modeling p…

Thanks. I have been recently reading about the integrating codes in Custom theano Op to do numerical integration . I try to embed the codes with the above model where: func = pm.math.exp(-(beta*(t - t0))**2) integrate = Integrate(func, t, beta, t0) integral_results= integrate(start, stop, beta, t…

Best way I would say is to use pm.Data, which is design to handle prediction better: training_set = df_res[~df_res.index.isin(t_pred_idx)] test_set = df_res.loc[t_pred_idx] with pm.Model() as model_logistic: predictor = pm.Data('predictor', training_set.t.values) observed = pm.Data('observe…

Thank you for your reply. In doing so, I got some errors: ValueError: Input dimension mis-match. (input[0].shape[0] = 3069, input[1].shape[0] = 33) The training_set dataset has 3069 rows, and 33 levels (levels being the provinces of china, similar to the radon example, in which case there are 85 c…

I see, you actually need to build a 2d mu and index both the provinces and timepoint. I am not sure if it makes the problem more complicated. Basically, you will need to get a single time series time index t as predictor, make sure that predictor-t0 broadcast correctly so that the shape is [number_p…

Thank you so much! But I still got some problems: Following your advice and previous adjustment to the codes, I rewrote the model, but now I am still stuck in the mis-match error: ValueError: Input dimension mis-match. (input[0].shape[1] = 93, input[1].shape[1] = 33) with pm.Model() as model_log…

Try this: n_province, n_timepoint = 33, 93 # replace below with your real data predictor_val = np.repeat( np.linspace(0., 1., n_timepoint)[np.newaxis, ...], n_province, axis=0) # want shape = (33, 93) observed_val = np.random.randn(n_province, n_timepoint) with pm.Model() as model_log…

Thank you again! Your codes work correctly! However, when also consider the prediction problem, I assemble the following codes, which prompted errors indicating shapes not correct for the predictor dataset? ValueError: Input dimension mis-match. (input[0].shape[1] = 25, input[1].shape[1] = 93) Do …

No need to do indexing: n_timepoint_forecast = 20 predictor_forecast = np.repeat( np.linspace(1., 1. + 1./n_timepoint*n_timepoint_forecast, n_timepoint_forecast)[np.newaxis, ...], n_province, axis=0) with model_logistic: pm.set_data({'predictor': predictor_forecast}) pm.sample_…

Finally! It works! Amazing! Thank you! But, I am still wondering, if the above codes work for 2-d mu, then, if I have more dimensional data, it would also work right? For example, given the tuple with a shape like: (province_number, city_number, time_poins), I just need to change shape in alpha (an…

depends, since city is nested within province, you can probably model it with (city_number, time_poins) - however, if you have another predictor that is complete orthogonal to the other dimension then you can do what you describing.

Using Pymc3 to do forecasting and numerical integration

Questions

Antonio_Lam May 5, 2020, 3:21pm 12

Found great solution here: PyMC3 : Using a parameter as a limit of integral

Topic		Replies	Views
PyMC3 : Using a parameter as a limit of integral Questions	6	2893	July 1, 2018
Getting the current sample variable value so it can be used to compute the observed inputs in a secondary model likelihood Questions	10	2307	June 25, 2018
Implementing a more sophisticated pymc3 model v3 theano , modeling	8	829	June 11, 2022
Regression model sampling solely one sample every 5th second version agnostic modeling	2	540	August 29, 2022
Prediction using sample_ppc in Hierarchical model Questions from_github	6	5022	December 14, 2017

Using Pymc3 to do forecasting and numerical integration

Related topics